The Stata Journal (2015)

15, Number 2, pp. 480500

Conducting interrupted time-series analysis for

single- and multiple-group comparisons
Ariel Linden
Linden Consulting Group, LLC
Ann Arbor, MI
alinden@lindenconsulting.org

Abstract. In this article, I introduce the itsa command, which performs in-
terrupted time-series analysis for single- and multiple-group comparisons. In an
interrupted time-series analysis, an outcome variable is observed over multiple,
equally spaced time periods before and after the introduction of an intervention
that is expected to interrupt its level or trend. The itsa command estimates the
eect of an intervention on an outcome variable either for a single treatment group
or when compared with one or more control groups. Additionally, its options al-
low the user to control for autocorrelated disturbances and to estimate treatment
eects over multiple periods.
Keywords: st0389, itsa, interrupted time series, quasi-experimental designs, causal
inference

1 Introduction
In considering the impact of large-scale interventions (for example, population-based
health interventions, media campaigns, and dissemination of professional guidelines)
or public policy changes (for example, new laws or taxes), researchers are often faced
with an eective sample size of N = 1, where the treated group may be the local
community, state, or an even larger unit. It is also fairly common in these situations
that the only data available are reported at an aggregate level (for example, morbidity
or mortality rates, average costs, and median incomes). If multiple observations on an
outcome variable of interest in the preintervention and postintervention periods can be
obtained, an interrupted time-series analysis (ITSA) oers a quasi-experimental research
design with a potentially high degree of internal validity (Campbell and Stanley 1966;
Shadish, Cook, and Campbell 2002). Naturally, when the treated groups outcomes can
also be contrasted with those of one or more comparison groups, the internal validity
is further enhanced by allowing the researcher to potentially control for confounding
omitted variables.
ITSA has been used in many areas of study, such as assessing the eects of commu-
nity interventions (Biglan, Ary, and Wagenaar 2000; Gillings, Makuc, and Siegel 1981),
public policy (Muller 2004), regulatory actions (Briesacher et al. 2013), and health tech-
nology assessment (Ramsay et al. 2003), to name but a few. ITSA has also been proposed
as a more exible and rapid design to be considered in health research before defaulting
to the traditional two-arm randomized controlled trial (Riley et al. 2013). In addition,

2015 StataCorp LP st0389

A. Linden 481

systematic reviews of the literature are increasingly including studies that have used
ITSA as their primary research design (Cochrane Eective Practice and Organisation of
Care [EPOC] 2013).
In this article, I introduce the new itsa command, which performs interrupted
time-series analysis using two ordinary least-squares (OLS) regression-based approaches
available in the ocial Stata packages newey and prais. Additionally, itsa can estimate
treatment eects for multiple treatment periods.

2 Method and formulas

Statistical analyses used for ITSA must account for autocorrelated data. The two general
approaches historically used in ITSA are autoregressive integrated moving-average mod-
els (see Box and Tiao [1975], Glass, Willson, and Gottman [1975], and McDowall et al.
[1980]) and OLS regression models designed to adjust for autocorrelation (see, among
others, Crosbie [1993]; Gottman [1981]; McKnight, McKean, and Huitema [2000]; Si-
monton [1977a]; and Velicer and McDonald [1991]). itsa relies on OLS rather than on
regression methods based on autoregressive integrated moving-average models because
the former is often more exible and broadly applicable in an interrupted time-series
context (Box and Jenkins 1976; Velicer and Harrop 1983).

2.1 The single-group analysis

When there is only one group under study (no comparison groups), the standard ITSA
regression model assumes the following form (Huitema and McKean 2000a; Linden and
Adams 2011; Simonton 1977a; Simonton 1977b):

Y t = 0 + 1 Tt + 2 Xt + 3 Xt Tt +  t (1)

Yt is the aggregated outcome variable measured at each equally spaced time point t, Tt is
the time since the start of the study, Xt is a dummy (indicator) variable representing the
intervention (preintervention periods 0, otherwise 1), and Xt Tt is an interaction term.
These terms are displayed in the lower half of gure 1. In the case of a single-group
study, 0 represents the intercept or starting level of the outcome variable. 1 is the
slope or trajectory of the outcome variable until the introduction of the intervention.
2 represents the change in the level of the outcome that occurs in the period immedi-
ately following the introduction of the intervention (compared with the counterfactual).
3 represents the dierence between preintervention and postintervention slopes of the
outcome. Thus we look for signicant p-values in 2 to indicate an immediate treatment
eect, or in 3 to indicate a treatment eect over time (Linden and Adams 2011).
482 Conducting interrupted time-series analysis

Figure 1. Visual depiction of a single group (lower line) and multiple group (upper and
lower lines) interrupted time-series design, from Linden and Adams (2011)

Legend:
Single group0 : intercept; 1 : slope prior to intervention; 2 : change in level in the
period immediately following intervention initiation (compared with counterfactual);
3 : dierence between preintervention and postintervention slopes.

Multiple group0 to 3 represent the control group; 4 to 7 represent the

treatment group. 4 : dierence in the level between treatment and control prior
to intervention; 5 : dierence in the slope between treatment and control prior to
intervention; 6 : dierence in the level between treatment and control in the period
immediately following intervention initiation; 7 : dierence between treatment and
control in the slope after initiation of the intervention compared with preintervention.

When the random error terms follow a rst-order autoregressive [AR(1)] process,

t = t1 + ut (2)

where the autocorrelation parameter is the correlation coecient between adjacent

error terms, such that || < 1, and the disturbances ut are independent N (0, 2 ) (see
Kutner et al. [2005] for a detailed discussion of autocorrelation in time-series models).
Identication in both the single- and multiple-group models is driven by the func-
tional-form assumptions of the ITSA model. By design, a single-group ITSA has no
comparable control group; rather, the preintervention trend projected into the treat-
ment period serves as the counterfactual. We assume that any time-varying unmeasured
A. Linden 483

confounder is relatively slowly changing so that it would be distinguishable from the

sharp jump of the intervention indicator. This underscores the need for caution with
these methods if there are multiple policy shifts in the time window around the imple-
mentation of the intervention.
The assumptions necessary for causal inference in the single-group ITSA may seem
plausible when the preintervention trend is at followed by a signicant change in the
outcome variable immediately following the introduction of the intervention and then
sustained over time. However, these assumptions may seem less plausible if a trend
already exists in the time series prior to the intervention. While the ITSA literature
does not address the topic of testing for interruptions in the level and trend of the
outcome variable [2 and 3 of (1)] prior to the actual period in which the intervention
started, we can look to the regression-discontinuity literature to provide guidance for
applicable robustness tests. In practice, this would simply entail testing for interruptions
after replacing the true intervention start period with other pseudo-start periods along
the preintervention continuum.
In an adaptation of Imbens and Lemieux (2008) for ITSA, an investigator could use
the median time point of the preintervention period to test for an interruption. In a
suciently long time series, the median time point of the preintervention period is a
good choice of a pseudo-start period to maximize power to detect a signicant jump
(because the subsample will be evenly split on both sides). For shorter time series,
a simple iterative process of testing each preintervention time period as the pseudo-
start period may be a good approach. In using such robustness tests, the underlying
assumptions of the single-group ITSA may be challenged if interruptions in the level or
trend of the outcome variable are found to exist at other time points prior to the true
initiation of the intervention.

2.2 The multiple-group analysis

When one or more control groups are available for comparison, the regression model in
(1) is expanded to include four additional terms (4 to 7 ) (Linden and Adams 2011;
Simonton 1977a; Simonton 1977b):
Yt = 0 + 1 Tt + 2 Xt + 3 Xt Tt + 4 Z + 5 ZTt + 6 ZXt + 7 ZXt Tt + t (3)
Here Z is a dummy variable to denote the cohort assignment (treatment or control), and
ZTt , ZXt , and ZXt Tt are all interaction terms among previously described variables.
Now when examining gure 1, the coecients of the lower line, 0 to 3 , represent the
control group, and the coecients of the upper line, 4 to 7 , represent values of the
treatment group. More specically, 4 represents the dierence in the level (intercept) of
the outcome variable between treatment and controls prior to the intervention, 5 rep-
resents the dierence in the slope (trend) of the outcome variable between treatment
and controls prior to the intervention, 6 indicates the dierence between treatment and
control groups in the level of the outcome variable immediately following introduction of
the intervention, and 7 represents the dierence between treatment and control groups
in the slope (trend) of the outcome variable after initiation of the intervention compared
484 Conducting interrupted time-series analysis

with preintervention (akin to a dierence-in-dierences of slopes). If the multiple-group

model follows an AR(1) process, the random error term is dened as in (2).
A multiple-group ITSA may be particularly valuable when there is an exogenous
policy shift that aects all the groups. The key assumption is that the change in
the level or trend in the outcome variable is presumed to be the same both for the
control group and, counterfactually, for the treatment group had it not received the
intervention. In other words, we assume that confounding omitted variables aect both
treatment and control groups similarly. A major strength of the multiple-group ITSA
is the ability to test for comparability between groups on observed covariates and in
particular, the two parameters 4 and 5 , which play a particularly important role in
establishing whether the treatment and control groups are balanced on both the level
and the trajectory of the outcome variable in the preintervention period. If these data
were from a randomized controlled trial, we would expect similar levels and slopes prior
to the intervention. However, in an observational study where equivalence between
groups cannot be ensured, any observed dierences will likely raise concerns about the
ability to draw causal inferences about the relationship between the intervention and
the outcomes (Linden and Adams 2011).
To reduce the threat of confounding, investigators may attempt to emulate the
randomization process with observational data by nding control groups that are com-
parable to the treatment group on observed preintervention covariates. One approach
to nding comparable controls out of a pool of potential candidates is via an iterative
process in which each non-treated group is compared separately with the treatment
group using the model dened in (3). Those groups who have p-values greater than 0.05
(or a higher threshold) on both 4 and 5 can be selected as controls for inclusion in the
nal model. This method can be easily expanded to other available covariates; however,
there is a diminished likelihood of nding good controls as the number of covariates is in-
creased. If achieving balance on many covariates is an important factor, two alternative
approaches to itsa should be considered: the synthetic controls approach described by
Abadie, Diamond, and Hainmueller (2010) and implemented in Stata using the synth
package (Abadie, Diamond, and Hainmueller 2014), or the propensity-score weighting
technique described by Linden and Adams (2011).

2.3 Data variables corresponding to model parameters

Table 1 displays the variables used in regression models (1) and (3), using an articial ex-
ample with one intervention period. There are two individuals (or groups) in these data
(ID = 1, 2) with six observations each (T ). X indicates that there are two preintervention
observations, followed by four observations in the intervention period (the intervention
commences when T = 3). XT is an interaction term of X T , which starts in the
observation period immediately following the start of the intervention (T = 4) and runs
sequentially until the last observation when T = 6 (see Huitema and McKean [2000a]
for an exposition on the appropriateness of commencing the sequence in the observation
period after the start of the intervention). Here we transform XT = (T 3) X so that
it runs sequentially starting at 1. Additional variables are required for a multiple-group
A. Linden 485

analysis. Z indicates the treatment status, where Z = 1 for the treatment group and
Z = 0 for the control group. ZT , ZX, and ZXT are additional interaction terms used
in multiple-group comparisons, as described above in section 2.2. When multiple treat-
ment periods are specied, additional variables are added to the dataset, corresponding
to each treatment period respectively (see section 4.3 for an example).
Upon running the itsa command, all variables required for the corresponding single-
or multiple-group model are automatically generated and added to the dataset.

Table 1. Covariates used in a single-group ITSA (T , X, XT ) and multiple-group ITSA (T ,

X, XT , Z, ZT , ZX, ZXT ) corresponding to regression models (1) and (3), respectively

ID T X XT Z ZT ZX ZXT
1 1 0 0 1 1 0 0
1 2 0 0 1 2 0 0
1 3 1 0 1 3 1 0
1 4 1 1 1 4 1 1
1 5 1 2 1 5 1 2
1 6 1 3 1 6 1 3

2 1 0 0 0 0 0 0
2 2 0 0 0 0 0 0
2 3 1 0 0 0 0 0
2 4 1 1 0 0 0 0
2 5 1 2 0 0 0 0
2 6 1 3 0 0 0 0

2.4 Models
itsa allows the user to choose between two OLS regressionbased models specically de-
signed for time-series data. The rst, newey (see [TS] newey), estimates the coecients
by OLS regression but produces NeweyWest standard errors to handle autocorrelation
in addition to possible heteroskedasticity. The second model, prais (see [TS] prais),
uses the generalized least-squares method to estimate the parameters in a linear regres-
sion model in which the errors are assumed to follow an AR(1) process. More specically,
prais oers several methods to transform the original observations based on the pooled
autocorrelation estimate p to remove the correlation between rst-order errors (that is,
the correlation between the errors of each observation period and those of the preceding
observation period).
486 Conducting interrupted time-series analysis

The type of model that an investigator will choose for conducting time-series analysis
will likely depend on a combination of factors, with primary attention on the number of
lags in the data for which autocorrelation is present. In general, the investigator rst ts
an OLS model using either regress or newey (with lag(0) specied) and then tests for
autocorrelation in the error distribution. It is important to test for the presence of au-
tocorrelated errors when using regression-based time-series methods, because such tests
provide critical diagnostic information regarding the adequacy of the time-series model
(that is, whether tests and condence intervals on the regression coecients are satisfac-
tory, whether important variables have been left out of the time-series regression model
and because autocorrelated errors are produced when the functional form of the vari-
ables included in the model is incorrect; Huitema and McKean [2000b]). The package
oers several postestimation commands for this purpose (see [R] regress postestima-
tion time series). In addition, there is a comprehensive and versatile user-written
program, actest (Baum and Schaer 2013), that is downloadable from the Statistical
Software Components archive, with the default being the CumbyHuizinga general test
for autocorrelation (Cumby and Huizinga 1992).

3 The itsa command

3.1 Syntax
        
itsa depvar indepvars if in weight , trperiod(numlist) single
treatid(#) contid(numlist) prais lag(#) figure posttrend replace

prefix(string) model options

A dataset for a single panel must be declared to be time-series data using tsset timevar.
When the dataset contains multiple panels, a strongly balanced panel dataset using
tsset panelvar timevar must be declared; see [TS] tsset. indepvars may contain fac-
tor variables; see [U] 11.4.3 Factor variables. depvar and indepvars may contain
time-series operators; see [U] 11.4.4 Time-series varlists. aweights are allowed; see
[U] 11.1.6 weight. See [TS] newey postestimation and [TS] prais postestimation
for features available after estimation.

3.2 Options
trperiod(numlist) species the time period when the intervention begins. The values
entered for the time period must be in the same units as the panel time variable
specied in tsset timevar; see [TS] tsset. More than one period may be specied.
trperiod() is required.
single indicates that itsa will be used for a single-group analysis. Conversely, omitting
single indicates that itsa is for a multiple-group comparison.
A. Linden 487

treatid(#) species the identier of the single treated unit under study when the
dataset contains multiple panels. The value entered must be in the same units as
the panel variable specied in tsset panelvar timevar; see [TS] tsset. When the
dataset contains data for only a single panel, treatid() must be omitted.
contid(numlist) species a list of identiers to be used as control units in the multiple-
group analysis. The values entered must be in the same units as the panel variable
specied in tsset panelvar timevar; see [TS] tsset. If contid() is not specied, all
nontreated units in the data will be used as controls.
prais species to t a prais model. If prais is not specied, itsa will use newey as
the default model.
lag(#) species the maximum lag to be considered in the autocorrelation structure
when a newey model is chosen. If the user species lag(0), the default, the output
is the same as regress, vce(robust). An error message will appear if both prais
and lag() are specied, because prais implements an AR(1) model by design.
figure produces a line plot of the predicted depvar variable combined with a scatterplot
of the actual values of depvar over time. In a multiple-group analysis, figure plots
the average values of all controls used in the analysis (more specically, data for
specied controls are collapsed and the monthly observations are averaged).
posttrend produces posttreatment trend estimates using lincom, for the specied
model. In the case of a single-group ITSA, one estimate is produced. In the case of
a multiple-group ITSA, an estimate is produced for the treatment group, the control
group, and the dierence. In the case of multiple treatment periods, a separate table
is produced for each treatment period.
replace replaces variables created by itsa if they already exist. If prefix() is specied,
only variables created by itsa with the same prex will be replaced.
prefix(string) adds a prex to the names of variables created by itsa. Short prexes
are recommended.
model options specify all available options for prais when the prais option is chosen;
otherwise, all available options for newey other than lag() are specied.

3.3 Stored results

Because itsa passes all user-entered information to prais and newey, all results stored
by those commands are available. Additionally, itsa generates several key time-series
variables and adds them to the current dataset, as described in section 2.3. These
additional variables allow the user to further estimate treatment eects using arima or
other time-series models.
Table 2 is a cross reference to default names for those variables that appear in the
regression output tables (and used when posttrend is specied). Variables starting
with z are added to the dataset only when a multiple-group comparison is specied.
488 Conducting interrupted time-series analysis

(trperiod) is a sux added to certain variables indicating the start of the intervention
period. This is particularly helpful for dierentiating between added variables when
multiple interventions are specied (see the example presented in section 4.3). If the
user species a prefix(), it will be applied to all variables generated by itsa.

Table 2. Descriptions of default names for variables that appear in the regression output
tables

Variable Description
t time since start of study
x(trperiod) dummy variable representing the intervention periods
(preintervention periods 0, otherwise 1)
x t(trperiod) interaction of x and t
z dummy variable to denote the cohort assignment
(treatment or control)
z x(trperiod) interaction of z and x
z x t(trperiod) interaction of z, x, and t
s depvar pred predicted value generated after running itsa for a
single group
m depvar pred predicted value generated after running itsa for a
multiple-group comparison

4 Examples
In 1988, California passed the voter-initiative Proposition 99, which was a widespread
eort to reduce smoking rates by raising the cigarette excise tax by 25 cents per pack and
to fund anti-smoking campaigns and other related activities throughout the state (for
a comprehensive discussion of this initiative, see Abadie, Diamond, and Hainmueller
[2010]). Per capita cigarette sales (in packs) is the most widely used indicator of
smoking prevalence found in the tobacco research literature (Abadie, Diamond, and
Hainmueller 2010) and serves here as the aggregate outcome variable under study,
measured at the state level from 1970 until 2000 (with 1989 representing the rst
year of the intervention). The current data le was obtained from the synth pack-
age (Abadie, Diamond, and Hainmueller 2014), which originally obtained the cigarette
sales data and average retail price of cigarettes from Orzechowski and Walker (2005).
Eleven states were discarded from the dataset because of their adoption of some other
large-scale tobacco control program at some point during Californias intervention pe-
riod under study between 1989 and 2000, leaving 38 states as potential controls (Abadie,
Diamond, and Hainmueller 2010).
A. Linden 489

4.1 Single-group ITSA

In this example, we use itsa to assess the impact of Proposition 99 in reducing Califor-
nias per capita cigarette sales (in packs), using a single-group design. More specically,
we assess whether the introduction of Proposition 99 resulted in a shift in the level and
trend of per capita cigarette sales compared with those of the preintervention period
(as described in section 2.1).
First, we load the data and declare the dataset as panel:

. use cigsales
. tsset state year
panel variable: state (strongly balanced)
time variable: year, 1970 to 2000
delta: 1 unit

Next, we specify a single-group ITSA with California (state number 3 in the study) as
the treatment group and 1989 as the start of the intervention, request postintervention
trend estimates, and plot the results. The model is estimated using newey with one lag:

. itsa cigsale, single treat(3) trperiod(1989) lag(1) posttrend figure

panel variable: state (strongly balanced)
time variable: year, 1970 to 2000
delta: 1 unit
Regression with Newey-West standard errors Number of obs = 31
maximum lag: 1 F( 3, 27) = 331.45
Prob > F = 0.0000

Newey-West
cigsale Coef. Std. Err. t P>|t| [95% Conf. Interval]

_t -1.779474 .3834188 -4.64 0.000 -2.566184 -.9927632

_x1989 -20.0581 4.724395 -4.25 0.000 -29.75175 -10.36444
_x_t1989 -1.494652 .4368201 -3.42 0.002 -2.390933 -.5983715
_cons 134.0053 4.600271 29.13 0.000 124.5663 143.4442

Postintervention Linear Trend: 1989

Treated: _b[_t]+_b[_x_t1989]

Linear Trend Coeff Std. Err. t P>|t| [95% Conf. Interval]

Treated -3.2741 0.2688 -12.1803 0.0000 -3.8257 -2.7226

As shown in the regression table, the starting level of the per capita cigarette sales
was estimated at 134 packs, and sales appeared to decrease signicantly every year prior
to 1989 by 1.78 packs (P < 0.0001, CI = [2.57, 0.99]). In the rst year of the inter-
vention (1989), there appeared to be a signicant decrease in per capita cigarette sales
of 20.06 packs (P < 0.0001, CI = [29.75, 10.36]), followed by a signicant decrease in
the annual trend of sales (relative to the preintervention trend) of 1.49 packs per capita
per year (P = 0.002, CI = [2.39, 0.60]). We also see, from the lincom estimate pro-
duced by specifying posttrend, that after the introduction of Proposition 99, per capita
490 Conducting interrupted time-series analysis

cigarette sales decreased annually at a rate of 3.27 packs (95% CI = [3.83, 2.72]). Fig-
ure 2 provides a visual display of these results.

California
100 120 140 Intervention starts: 1989
Cigarette sales percapita (in packs)
40 60 80

1970 1980 1990 2000

Year

Actual Predicted
Regression with NeweyWest standard errors lag(1)

Figure 2. Single-group ITSA with NeweyWest standard errors and one lag

To ensure that we t a model that accounts for the correct autocorrelation structure,
we use actest (Baum and Schaer 2013), to test for autocorrelation.

. actest, lags(6)
Cumby-Huizinga test for autocorrelation
H0: variable is MA process up to order q
HA: serial correlation present at specified lags >q

H0: q=0 (serially uncorrelated) H0: q=specified lag-1

HA: s.c. present at range specified HA: s.c. present at lag specified

lags chi2 df p-val lag chi2 df p-val

1 - 1 15.242 1 0.0001 1 15.242 1 0.0001

1 - 2 15.255 2 0.0005 2 3.300 1 0.0693
1 - 3 15.325 3 0.0016 3 1.192 1 0.2749
1 - 4 15.896 4 0.0032 4 0.000 1 0.9880
1 - 5 16.057 5 0.0067 5 1.113 1 0.2914
1 - 6 16.078 6 0.0133 6 2.051 1 0.1521

Test allows predetermined regressors/instruments

Test requires conditional homoskedasticity
A. Linden 491

As shown in the right-side panel of the output table, autocorrelation is present at

lag 1 but not at any higher lag orders (up to the six lags tested). Thus our initial model
specifying lag(1) should correctly account for this autocorrelation.
An alternative approach is to rerun itsa specifying the prais option, which is
inherently designed to t an AR(1) model. Here we specify rhotype(tscorr), which
bases p on the autocorrelation of the residuals, and add robust standard errors.

. itsa cigsale, single treat(3) trperiod(1989) replace prais rhotype(tscorr)

> vce(robust)
panel variable: state (strongly balanced)
time variable: year, 1970 to 2000
delta: 1 unit
(output omitted )
Prais-Winsten AR(1) regression -- iterated estimates
Linear regression Number of obs = 31
F(3, 27) = 609.24
Prob > F = 0.0000
R-squared = 0.9011
Root MSE = 2.5964

Semirobust
cigsale Coef. Std. Err. t P>|t| [95% Conf. Interval]

_t -1.843139 .4538631 -4.06 0.000 -2.77439 -.9118892

_x1989 -6.094491 .8840197 -6.89 0.000 -7.90835 -4.280633
_x_t1989 -1.998494 .9191 -2.17 0.039 -3.884332 -.1126568
_cons 128.1931 3.958813 32.38 0.000 120.0703 136.316

rho .9424635

Durbin-Watson statistic (original) 0.535242

Durbin-Watson statistic (transformed) 1.342728

Because the estimates produced using prais are transformed, they are not directly
comparable with those of newey, which are produced using an OLS model. However,
these results conrm a signicant decrease in the annual trend of sales (relative to the
preintervention trend) of 2 packs per capita per year (P = 0.039, CI = [3.88, 0.11]).
prais provides the DurbinWatson d statistic as an indicator of how well the model
corrects for rst-order autocorrelation. d can take on values between 0 and 4, and under
the null hypothesis, d is equal to 2. Values of d less than 2 suggest positive autocor-
relation (p > 0), whereas values of d greater than 2 suggest negative autocorrelation
(p < 0); see [R] regress postestimation time series. As discussed previously, there
are several more intuitive and exible tests of autocorrelation; however, none of them
can currently be used in conjunction with prais.
492 Conducting interrupted time-series analysis

4.2 Multiple-group ITSA

In this example, we use itsa to assess the impact of Proposition 99 in reducing Cal-
ifornias per capita cigarette sales (in packs), using a multiple-group design. More
specically, we now compare Californias experience with that of the other 38 states in
the data le.

. itsa cigsale, treat(3) trperiod(1989) lag(1) replace figure

panel variable: state (strongly balanced)
time variable: year, 1970 to 2000
delta: 1 unit
Regression with Newey-West standard errors Number of obs = 1,209
maximum lag: 1 F( 7, 1201) = 364.04
Prob > F = 0.0000

Newey-West
cigsale Coef. Std. Err. t P>|t| [95% Conf. Interval]

_t -.5477701 .2941289 -1.86 0.063 -1.124834 .0292935

_z -2.041967 5.75639 -0.35 0.723 -13.33567 9.251731
_z_t -1.231704 .4641182 -2.65 0.008 -2.142276 -.321131
_x1989 -17.25168 3.815452 -4.52 0.000 -24.73737 -9.765987
_x_t1989 -.5035089 .5252893 -0.96 0.338 -1.534096 .5270778
_z_x1989 -2.806417 5.841839 -0.48 0.631 -14.26776 8.654929
_z_x_t1989 -.9911435 .6657528 -1.49 0.137 -2.297311 .3150244
_cons 136.0472 3.818559 35.63 0.000 128.5554 143.539

As shown in the regression table, the initial mean level dierence between California
and the remaining states ( z) was not signicant (P = 0.723, CI = [13.33, 9.25]),
but the dierence in the mean baseline slope ( z t) was signicant (P = 0.008, CI =
[2.14, 0.32]). This is veried upon visual inspection of gure 3: the trajectory of
mean cigarette sales for the 38 states appears to rise higher than in California, and that
level remains elevated throughout the duration of the observation period. Given this
dierential pattern of change in the baseline, one could argue that the 38 other states
were not comparable with California and, thus, treatment-eect estimates for z x1989
and z x t1989 may be biased (in the present case, both estimates are not statistically
signicant). Therefore, this model could be improved by limiting the choice of control
groups to only those with similar values on these two variables.
A. Linden 493

California and average of controls

Intervention starts: 1989

Cigarette sales percapita (in packs)

40 60 80 100 120 140

1970 1980 1990 2000

Year

California: Actual Predicted

Controls average: Actual Predicted
Regression with NeweyWest standard errors lag(1)

Figure 3. Multiple-group ITSA with NeweyWest standard errors and one lag; all 38
nontreated states are used for comparison

In the following example, we limit the analysis to only those states that are compa-
rable with California on baseline level and trend of the outcome variable, as described in
section 2.2. Comparability in the current context is dened as having a p-value greater
than 0.10 on both z and z t. Three comparison states meet this criteria: Colorado,
Idaho, and Montana.
As shown in both the regression table and veried upon visual inspection of gure 4,
the treatment group is comparable with controls on both baseline level and trend.
While there is no statistically signicant treatment eect during the rst year of the
intervention ( z x1989), there is a statistically signicant annual reduction in the pre
post trend compared with that of controls of 1.97 per capita cigarette sales per year
(P = 0.003, CI = [3.26, 0.68]). Additionally, we see from the posttrend output that
the treatment group decreased annual cigarette sales in the postintervention period by
3.27 packs, the control group decreased sales over the same period by only 1 pack, and
the dierence between them is 2.28 packs per capita per year.
494 Conducting interrupted time-series analysis

. itsa cigsale, treat(3) trperiod(1989) contid(4 8 19) lag(1) replace posttrend

> figure
panel variable: state (strongly balanced)
time variable: year, 1970 to 2000
delta: 1 unit
Regression with Newey-West standard errors Number of obs = 124
maximum lag: 1 F( 7, 116) = 251.48
Prob > F = 0.0000

Newey-West
cigsale Coef. Std. Err. t P>|t| [95% Conf. Interval]

_t -1.464503 .3837773 -3.82 0.000 -2.224622 -.7043836

_z 2.046198 6.218666 0.33 0.743 -10.27065 14.36305
_z_t -.3149707 .5330632 -0.59 0.556 -1.37077 .7408282
_x1989 -13.58866 4.180499 -3.25 0.002 -21.86867 -5.308658
_x_t1989 .4746428 .4992501 0.95 0.344 -.514185 1.463471
_z_x1989 -6.469433 6.185239 -1.05 0.298 -18.72008 5.781212
_z_x_t1989 -1.969295 .6533782 -3.01 0.003 -3.263393 -.6751973
_cons 131.9591 4.355318 30.30 0.000 123.3328 140.5853

Comparison of Linear Postintervention Trends: 1989

Treated : _b[_t] + _b[_z_t] + _b[_x_t1989] + _b[_z_x_t1989]
Controls : _b[_t] + _b[_x_t1989]
Difference : _b[_z_t] + _b[_z_x_t1989]

Linear Trend Coeff Std. Err. t P>|t| [95% Conf. Interval]

Treated -3.2741 0.2594 -12.6234 0.0000 -3.7878 -2.7604

Controls -0.9899 0.2883 -3.4336 0.0008 -1.5608 -0.4189

Difference -2.2843 0.3878 -5.8905 0.0000 -3.0523 -1.5162

These results highlight the importance of ensuring that treatment and control units
are comparable on the preintervention level and trend of the outcome variable when con-
ducting a multiple-group ITSA. As described in section 2.2, an iterative process can be
used in which each nontreated group is compared separately with the treatment group.
Those groups with p-values greater than a specied threshold on both 4 and 5 of 3 can
be retained as controls for inclusion in the nal model. This approach can be easily ex-
tended to other covariates as well; however, if achieving balance on many covariates is an
important factor, two alternative approaches to ITSA should be considered: the synthetic
controls approach described by Abadie, Diamond, and Hainmueller (2010) and imple-
mented in Stata using the synth package (Abadie, Diamond, and Hainmueller 2014),
or the propensity-score weighting technique described by Linden and Adams (2011).
A. Linden 495

California and average of controls

Intervention starts: 1989

Cigarette sales percapita (in packs)

40 60 80 100 120 140

1970 1980 1990 2000

Year

California: Actual Predicted

Controls average: Actual Predicted
Regression with NeweyWest standard errors lag(1)

Figure 4. Multiple-group ITSA with NeweyWest standard errors and one lag; three
states, comparable on the baseline level and trend of the outcome, are used for compar-
ison

4.3 Multiple treatment periods

itsa can accommodate design variations in which the eect of multiple treatment pe-
riods are of interest. For example, the researcher may be interested in studying the
eects of an intervention that is introduced, withdrawn, and reintroduced, or an in-
tervention that is followed by a separate intervention at a later point in time (see
Barlow, Hayes, and Nelson [1984] for many other design alternatives).
For exposition, in the following example we add a ctitious intervention to the
cigarette sales data, starting in 1982. We reestimate the single-group ITSA from sec-
tion 4.1, now with one additional intervention period.
496 Conducting interrupted time-series analysis

. itsa cigsale, single treat(3) trperiod(1982 1989) lag(1) replace posttr

> figure
panel variable: state (strongly balanced)
time variable: year, 1970 to 2000
delta: 1 unit
Regression with Newey-West standard errors Number of obs = 31
maximum lag: 1 F( 5, 25) = 657.58
Prob > F = 0.0000

Newey-West
cigsale Coef. Std. Err. t P>|t| [95% Conf. Interval]

_t -.2038463 .3343824 -0.61 0.548 -.8925198 .4848273

_x1982 -8.040472 2.875704 -2.80 0.010 -13.96309 -2.117849
_x_t1982 -3.639012 .378855 -9.61 0.000 -4.419278 -2.858745
_x1989 -9.285162 2.561198 -3.63 0.001 -14.56005 -4.010275
_x_t1989 .5687319 .3153918 1.80 0.083 -.0808296 1.218293
_cons 125.2333 2.281397 54.89 0.000 120.5347 129.932

Postintervention Linear Trend: 1982

Treated: _b[_t]+_b[_x_t1982]

Linear Trend Coeff Std. Err. t P>|t| [95% Conf. Interval]

Treated -3.8429 0.2237 -17.1757 0.0000 -4.3037 -3.3821

Postintervention Linear Trend: 1989

Treated: _b[_t]+_b[_x_t1982]+_b[_x_t1989]

Linear Trend Coeff Std. Err. t P>|t| [95% Conf. Interval]

Treated -3.2741 0.2793 -11.7205 0.0000 -3.8495 -2.6988

The interpretation of all coecients up to the second intervention is as before. That

is, the rst intervention period is compared with the preintervention period. However,
the additional coecients for the second intervention period, x1989 and x t1989, are
now compared with those of the prior (rst) intervention period.
As shown in both the regression table and veried upon visual inspection of gure 5,
there is evidence of a treatment eect beginning in 1982 and no additional decrease in
annual sales after the implementation of the second intervention (which in reality was
the true intervention period).
A. Linden 497

California
Intervention starts: 1982 1989

Cigarette sales percapita (in packs)

40 60 80 100 120

1970 1980 1990 2000

Year

Actual Predicted
Regression with NeweyWest standard errors lag(1)

Figure 5. Single-group ITSA with NeweyWest standard errors and two intervention
periods

We can demonstrate this eect further via the posttrend option, which estimates
the postintervention trends separately after the rst and second intervention periods.
As shown in the posttrend output, the annual decrease in cigarette sales after 1982 was
3.84 packs per year, while the annual decrease in sales after 1989 was slightly less, at
3.27 packs per year (the dierence is 0.569, which appears in the original regression table
as x t1989). Thus the results of this exercise reveal an additional utility of dening
multiple treatment periods in itsa when analyzing single-group data: it allows for the
testing of interruptions away from the true start of the intervention, as described in
section 2.1. As demonstrated in this example, a statistically signicant reduction in
the trend of cigarette sales in California began several years prior to implementation
of Proposition 99. This result also further highlights the importance, even when using
ITSA, of nding a comparable control group to represent the counterfactual.

5 Discussion
While the randomized controlled trial remains the gold standard research design, there
are situations in which this design is not feasible or practical, such as when large-scale
interventions or policy changes target the entire population. When data are available for
multiple time points in both the preintervention and the postintervention periods, inter-
rupted time-series designs oer a robust quasi-experimental alternative for evaluating
treatment eects (Campbell and Stanley 1966; Shadish, Cook, and Campbell 2002).
498 Conducting interrupted time-series analysis

In this article, I have demonstrated the basic implementation of itsa to estimate

treatment eects for a single treatment group, a multiple-group comparison, and when
more than one intervention has been employed sequentially. Additional important issues
were also addressed, such as criteria for choosing and specifying a model, testing for
autocorrelation, robustness testing for interruptions prior to the true intervention start
period, and choosing comparable controls. More-complex models can easily be estimated
with itsa by including additional covariates to control for confounding, seasonal eects,
and the impact of external events. Moreover, the addition of key time-series variables
to the dataset after running itsa allows for further estimation of treatment eects
using more complex OLS models or arima, assuming that more sophisticated time-
series modeling is warranted and assuming the availability of a sucient number of
observations.

6 Acknowledgments
I thank Nicholas J. Cox for his support while developing itsa, and I thank Steven J.
Samuels for creating the posttrend option and for assistance with several post-review
changes to itsa. I also thank Michael J. Harvey for his assistance with LATEX, and
Steven J. Samuels, Roger B. Newson, John L. Adams, Andrew Ryan, and Julia Adler-
Milstein for their reviews and helpful comments on the article. I also thank the anony-
mous reviewer and chief editor for their thoughtful reviews and recommendations for
improving both the article and the itsa command.

7 References
Abadie, A., A. Diamond, and J. Hainmueller. 2010. Synthetic control methods for com-
parative case studies: Estimating the eect of Californias tobacco control program.
Journal of the American Statistical Association 105: 493505.
. 2014. synth: Stata module to implement synthetic control methods for com-
parative case studies. Statistical Software Components S457334, Department of Eco-
nomics, Boston College. https://ideas.repec.org/c/boc/bocode/s457334.html.
Barlow, D. H., S. C. Hayes, and R. O. Nelson. 1984. The Scientist Practitioner: Research
and Accountability in Clinical and Educational Settings. New York: Pergamon Press.
Baum, C. F., and M. E. Schaer. 2013. actest: Stata module to perform Cumby
Huizinga general test for autocorrelation in time series. Statistical Software Compo-
nents S457668, Department of Economics, Boston College.
https://ideas.repec.org/c/boc/bocode/s457668.html.
Biglan, A., D. Ary, and A. C. Wagenaar. 2000. The value of interrupted time-series
experiments for community intervention research. Prevention Science 1: 3149.
Box, G. E. P., and G. M. Jenkins. 1976. Time Series Analysis: Forecasting and Control.
San Francisco, CA: Holden Day.
A. Linden 499

Box, G. E. P., and G. C. Tiao. 1975. Intervention analysis with applications to economic
and environmental problems. Journal of the American Statistical Association 70: 70
79.

Briesacher, B. A., S. B. Soumerai, F. Zhang, S. Toh, S. E. Andrade, J. L. Wagner,

A. Shoaibi, and J. H. Gurwitz. 2013. A critical review of methods to evaluate the
impact of FDA regulatory actions. Pharmacoepidemiology and Drug Safety 22: 986
994.

Campbell, D. T., and J. C. Stanley. 1966. Experimental and Quasi-Experimental De-

signs for Research. Chicago, IL: Rand McNally.

Cochrane Eective Practice and Organisation of Care (EPOC) Group. 2013. EPOC-
specic resources for review authors. Oslo: Norwegian Knowledge Centre for the
Health Services.
http://epocoslo.cochrane.org/epoc-specic-resources-review-authors.

Crosbie, J. 1993. Interrupted time-series analysis with brief single-subject data. Journal
of Consulting and Clinical Psychology 61: 966974.

Cumby, R. E., and J. Huizinga. 1992. Testing the autocorrelation structure of distur-
bances in ordinary least squares and instrumental variables regressions. Econometrica
60: 185195.

Gillings, D., D. Makuc, and E. Siegel. 1981. Analysis of interrupted time series mortality
trends: An example to evaluate regionalized perinatal care. American Journal of
Public Health 71: 3846.

Glass, G. V., V. L. Willson, and J. M. Gottman. 1975. Design and Analysis of Time-
Series Experiments. Boulder, CO: Colorado Associated University Press.

Gottman, J. M. 1981. Time-series Analysis: A Comprehensive Introduction for Social

Scientists. New York: Cambridge University Press.

Huitema, B. E., and J. W. McKean. 2000a. Design specication issues in time-series

intervention models. Educational and Psychological Measurement 60: 3858.

. 2000b. A simple and powerful test for autocorrelated errors in OLS intervention
models. Psychological Reports 87: 320.

Imbens, G. W., and T. Lemieux. 2008. Regression discontinuity designs: A guide to

practice. Journal of Econometrics 142: 615635.

Kutner, M. H., C. J. Nachtsheim, J. Neter, and W. Li. 2005. Applied Linear Statistical
Models. New York: McGrawHill.

Linden, A., and J. L. Adams. 2011. Applying a propensity-score based weighting model
to interrupted time series data: Improving causal inference in program evaluation.
Journal of Evaluation in Clinical Practice 17: 12311238.
500 Conducting interrupted time-series analysis

McDowall, D., R. McCleary, E. E. Meidinger, and R. A. Hay. 1980. Interrupted Time

Series Analysis. Newbury Park, CA: Sage Publications.

McKnight, S., J. W. McKean, and B. E. Huitema. 2000. A double bootstrap method

to analyze linear models with autoregressive error terms. Psychological Methods 5:
87101.

Muller, A. 2004. Floridas motorcycle helmet law repeal and fatality rates. American
Journal of Public Health 94: 556558.
Orzechowski, W., and R. C. Walker. 2005. The Tax Burden on Tobacco. Historical
Compilation, vol. 40. Arlington, VA: Orzechowski & Walker.
Ramsay, C. R., L. Matowe, R. Grilli, J. M. Grimshaw, and R. E. Thomas. 2003. In-
terrupted time series designs in health technology assessment: Lessons from two sys-
tematic reviews of behavior change strategies. International Journal of Technology
Assessment in Health Care 19: 613623.

Riley, W. T., R. E. Glasgow, L. Etheredge, and A. P. Abernethy. 2013. Rapid, respon-

sive, relevant (R3) research: A call for a rapid learning health research enterprise.
Clinical and Translational Medicine 2: 16.

Shadish, W. R., T. D. Cook, and D. T. Campbell. 2002. Experimental and Quasi-

Experimental Designs for Generalized Causal Inference. Boston: Houghton Miin.

Simonton, D. K. 1977a. Cross-sectional time-series experiments: Some suggested sta-

tistical analyses. Psychological Bulletin 84: 489502.
. 1977b. Erratum to Simonton. Psychological Bulletin 84: 1097.

Velicer, W. F., and J. Harrop. 1983. The reliability and accuracy of time series model
identication. Evaluation Review 7: 551560.

Velicer, W. F., and R. P. McDonald. 1991. Cross-sectional time series designs: A general
transformation approach. Multivariate Behavioral Research 26: 247254.

About the author

Ariel Linden is a health services researcher specializing in the evaluation of health care in-
terventions. He is both an independent consultant and an adjunct associate professor at the
University of Michigan in the department of Health Management and Policy, where he teaches
program evaluation.